Project: an agent with tools

The briefCopy Anchor Link

The user asks "Is this open-source project well maintained?" The agent calls tools for recent commits, open issues and the latest release, then writes a verdict with numbers.

StepsCopy Anchor Link

1. Define the tools with clear descriptions and schemas.
2. Write the loop by hand: call the model, run tool calls, append results, repeat until the model answers, with a step limit.
3. Validate every argument before running a tool.
4. When a tool fails, return an instructive error as the result.
5. Log every call and result.
6. Save the conversation so it can resume.

Check yourselfCopy Anchor Link

Run ten different projects, including one that does not exist and one with no recent activity. Read the logs. Count how often the agent picks the wrong tool or loops. Fix descriptions and limits, then rerun the same ten.

Done whenCopy Anchor Link

The agent answers five test questions using its tools. A repository that does not exist produces a recovered answer, not a crash. A step limit stops it when you force a tool to keep failing. Every tool call is logged with its arguments and result.

StretchCopy Anchor Link

Add a second tool for the project's documentation and let the agent decide which to use. Try a smaller model for the tool steps and compare cost and accuracy.

StarterCopy Anchor Link

Python. Reuse the loop from the agent loop lesson and give it two real tools backed by the GitHub API (no key needed for light use). Install pip install requests.

import requests

def recent_commits(repo: str) -> str:
    r = requests.get(f"https://api.github.com/repos/{repo}/commits", params={"per_page": 5}, timeout=10)
    if r.status_code == 404:
        return f"No repository named '{repo}'. Use the form owner/name, for example pallets/flask."
    r.raise_for_status()
    return "\n".join(f"{c['commit']['author']['date'][:10]}  {c['commit']['message'].splitlines()[0]}" for c in r.json())

def run_tool(name, args):
    if name == "recent_commits":
        return recent_commits(args["repo"])
    raise ValueError(f"unknown tool {name}")

Add a second tool for open issues and one for the latest release. Write the tool descriptions carefully: they are what the model sees.